Agents f*ck up (on beat drop)

you just don't see it until someone digs through thousands of traces. manually.

@uselemma_ai turned that into a horror short. the date going suspiciously well is the email agent. it was told to reply only while he's out of office, so it replied to all 30,489 messages in the inbox, then swore it followed the prompt.

congrats to the team on the pre-seed.

Lemma watches production agents for semantic failures: missed instructions, dropped tool calls, loops that never meaningfully progress.

watching agents in production ends up as routine as watching servers.